Welcome to Scribd. Sign in or start your free trial to enjoy unlimited e-books, audiobooks & documents.Find out more
Standard view
Full view
of .
Look up keyword
Like this
0 of .
Results for:
No results containing your search query
P. 1
Beyond Rules Getting More Out of Data With Belief Networks 2004 04

Beyond Rules Getting More Out of Data With Belief Networks 2004 04

Ratings: (0)|Views: 6|Likes:
Published by Joao

More info:

Published by: Joao on Jul 07, 2008
Copyright:Attribution Non-commercial


Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less





 © 2004 Charles River Analytics, All Rights Reserved
Beyond Rules:Getting More Out of Data with Belief Networks
Because Belief Networks greatly condense the information and relationships needed to create a decisionframework, they can be applied in situations previously too complex to be automated with rules-based expert anddecision-support systems.This paper describes Belief Networks, the next step beyond rules-based expert systems. It covers how Belief Networks can turn missing or uncertain data into valuable, actionable information. It starts by addressing therigidities and associated limitations of rule-based systems, and how Belief Networks condense the neededinformation and relationships for decision-making. The paper concludes with what you can do to start bringingBelief Networks to work for your organization.
If you are responsible for purchasing and implementing Expert Systems in your company
, read this paperto understand how Belief Networks can expand the strategic benefits of automated systems and decision supportsystems into areas that are too complex to benefit from business rules.
If you are a software vendor or consultant
, read this paper to understand how to incorporate Belief Networksinto your products and projects to offer new ways for your customers to gain the benefits of expert and decisionsupport systems in areas that are too complex to benefit from business rules.
Can you afford all the black-and-white rules you need…to cover the gray areas?
Rules-based expert systems use key insights to turn data into information and decisions. However, these systemshave a critical characteristic that restricts their use to only the simplest of problems: they only allow yes-or-no,if-this-then-that rules. When there are more than just a few contributing factors for a decision, or if some factorsinvolve gradations or uncertain information, the number of rules needed to fully define the decision makingprocess can grow exponentially. Additional rules are needed to account for situations where data is missing, orwhere inputs come in the form of possibilities or positions in a spectrum, not facts.
Do I really want to carry an umbrella every time I miss the weather report? 
Consider a simple system that would help you decide whether or not to carry an umbrella when you leave abuilding. You have several inputs you probably use in making this decision:
The most recent weather report you heard or saw
What the sky looks like (or looked like when last you saw it)
If you have achy joints, you may feel a twinge when the weather is about to changeThere are significant complexities and chances for uncertainty within even this small set of inputs. The weatherreporters, always hedging their bets, usually put a percentage on the chance of rain. You might miss the report if you’re in a hurry. Perhaps the weather isn’t changing fast enough for you to feel it in your achy joints. It maybe cloudy, but not so much as last time it rained…
Page 2 of 4© 2004 Charles River Analytics, All rights reservedTo understand how these additional complexities can cause the size of the system to grow unmanageably large,consider constructing a useful rule system for this situation. A number of simplifying assumptions will have tobe made to incorporate information that comes in the form of a spectrum (like the % chance of rain, or howcloudy it really is). In addition, you will have to manage a growing number of rules as you incorporate more of the important inputs and potential uncertainties.The weather report is easy to tackle:
If the chance of rain > 80%
Then always bring an umbrella
But, how do you know what to do if you miss the weather report?
If it is cloudy
Then always bring an umbrella?
Well, how do you define cloudy? It would seem extreme to bring an umbrella every time it was partly cloudy inthe morning. Is there a way to match this up with whether or not your joints hurt and what you might hear in theweather report?OK, let’s try to make some rules.
Aching joints have two states: either they hurt or they don’t
The weather report usually comes in some % chance of rain. Let’s break it into 4 ranges to make thingseasier: 0-40%, 41-70%, 71-90%, and 90-100%
Again, keeping things simple, let’s define 3 states for the sky: sunny, partly cloudy and cloudyNow combine these rules. That means…2 x 4 x 3 or 24 rules. You will need even more if you want to includecases like missing the weather report, or not being able see the sky because you’re indoors. Suddenly, the list of rules is getting long. Even if you decide to carry an umbrella no matter what if the weather report says chance of rain 90-100%, that only cuts out a few.
One Belief Network is Worth Thousands of Rules (or More)
In the case of the Umbrella Decision System, using a belief network instead of rules offers a chance to condensethe large number of required rules into a simple model. In fact, a belief network even allows the user toincorporate prior knowledge and uncertain information without adding rules or complexity.
The First Step: Create a Network and Generate a Table of Likelihoods 
A simple diagram of potentially useful inputs and the results they drive is the first step. You can now build atable that lays out which circumstances mean you should bring an umbrella.Notice that the weather report can be listed as either “will rain” or “will not rain” - we don’t need to split up thewhole range of possible % chances of rain because a belief network will allow us to enter a % chance directly.That is,
 you don’t have to be sure: the information that it
rain is just as useful to us as the information that it
Page 3 of 4© 2004 Charles River Analytics, All rights reserved
Squeezing all you can out of what you already know: 
If you can build a table like the one above for deciding to bring an umbrella, can’t you build similar ones foreach of the other inputs? You do know about how often it is cloudy, about how often the weather report calls forrain, and about how often your joints hurt. Can’t you somehow use this information to fill in the blanks whenyou happen to miss the weather report, or you are leaving for an evening meeting and can’t tell if it is cloudy?Yes. With a belief network, you can fill in for what you don’t know today with what you know usuallyhappened in the past.
Cloudy Partly Cloudy SunnyHistorical chances that it will be cloudy
30% 40% 30%
Usual Chance of Rain
How likely your joints are to hurt on any given day
Granted, this information about previous experience is not as good as knowing exactly what is going on rightnow, but
it is a lot better than using no information at all
to make a decision.
Other advantages of Belief Networks
Beyond the reduction in the expense of building a decision or expert system, Belief Networks offer a number of other advantages:
It is easy to explain how a system arrived at a particular recommendation, decision, or action
It is possible to diagnose problems: Belief Networks can be run in multiple directions
With sensitivity analysis, it is easy to understand how changing inputs can affect overall results
Where can I apply Belief Networks?
Belief networks have already been applied in a number of business areas. In each of the examples below, thesystem described would have been impossible with business rules.
Selecting loan opportunities with low default risk 
Estimating when equipment will need service, parts or retirement based on sensor input and records
Quickly resolving customer problems via online troubleshooting tools
Diagnosing and troubleshooting on-site cellular networks in real-time
Filtering email to accurately control spam and to highlight critical messages
Controlling autonomous vehicles and their navigation (wheeled, aquatic, aerospace, and others)
Assessing risks and rewards of R&D projects to move resources into the best opportunities
Many more…

You're Reading a Free Preview

/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->